Cross-Species Exposure-Response Meta-Analysis

Why Analyze Exposure-Response Across Species?

An exposure-response curve fitted in one species does not, by itself, establish the curve expected in humans. Species can differ in potency and in the steepness of the response, even when exposure and response are defined comparably. Cross-species exposure-response meta-analysis makes those differences explicit: it combines information across species while retaining species-level variation rather than treating all observations as if they came from one homogeneous population.

The unit of synthesis is often a species-specific exposure-response estimate, such as log EC50 and log Hill coefficient, together with its estimation uncertainty. The meta-analysis then asks how these estimates vary across species and whether measured species characteristics help explain that variation. A human projection is conditional on those modeling choices; it is not simply an average of animal responses.

Exposure-Response Parameters and the Hierarchical Model

For species \(s\), a common increasing Emax curve for response at exposure \(C\) is

Here \(E_{0,s}\) is baseline response, \(E_{\max,s}\) is the maximum change from baseline, \(EC_{50,s}\) is the exposure producing half of that change, and \(h_s\) is the Hill coefficient. For an inhibitory endpoint, the response direction or sign of the maximum change must be defined consistently. The potency parameter in this curve is \(EC_{50,s}\): a larger EC50 means lower potency.

This meta-analysis focuses on variation in potency and slope. Let \(\psi_s=(\log EC_{50,s},\log h_s)^\mathsf{T}\). Logging ensures positive EC50 and Hill coefficients when parameters are transformed back to their original scale. If species-specific fits also estimate baseline or maximum response, those parameters can be modeled in the hierarchy as well, but they should not be assumed common without justification.

The first equation represents the approximate sampling distribution of the species-specific estimates; \(V_s\) is their estimated covariance matrix. The second is the meta-regression: \(X_s\) contains an intercept and any species-level predictors, \(\beta\) contains their population coefficients, and \(b_s\) is the species-level deviation. The covariance matrix \(\Omega\) describes residual heterogeneity in log EC50 and log Hill coefficient after accounting for predictors. Its off-diagonal element allows species with higher-than-predicted EC50 (lower potency) to have systematically higher or lower-than-predicted Hill coefficients.

The approximation requires sufficiently informative species-specific data for the curve parameters to be estimable, and the reported \(V_s\) must represent their uncertainty, including covariance. Species should measure comparable endpoints at comparable effect times, and exposure should be defined on a biologically meaningful common scale. Conditional on the predictors and the model, species-level deviations are usually assumed independent and drawn from a common distribution. These assumptions are substantive: a hierarchical model cannot make incomparable assays or exposures equivalent simply by pooling them.

Combining Species Estimates

For one parameter without predictors, the random-effects model is \(\hat\theta_s\mid\theta_s\) approximately normal with variance \(v_s\), and \(\theta_s\) is normally distributed about a cross-species mean \(\mu_\theta\) with variance \(\tau^2\). Marginally, the estimate has variance \(v_s+\tau^2\). Thus a more precise species estimate receives more weight, but residual between-species variation prevents the most precise species from dominating as strongly as it could under a fixed-effect model.

If \(\tau^2\) is treated as known, the inverse-variance estimate of the mean is a weighted average. Its standard error follows from the variance of a sum of independent estimates. In practice, \(\tau^2\) is estimated, so uncertainty in that estimate can matter, especially when few species are available.

The last standard error is conditional on the specified heterogeneity variance and the normal approximation. It describes uncertainty in the cross-species mean, not the spread of true parameters in individual species. A new-species projection must also account for heterogeneity.

Meta-Regression and Human Projection

Pooling alone assumes that species are exchangeable around a common mean. If biological or assay characteristics plausibly explain some differences, a meta-regression can include them in \(X_s\). Examples might include a mechanistically motivated measure of target similarity or a relevant species-level physiological feature. Such predictors need a consistent definition and adequate variation across the included species. With a covariate, the model estimates the parameter relationship at a specified covariate value rather than assuming one universal value.

For a human projection, insert the human predictor values into the fitted relationship. The resulting \(X_H\hat\beta\) estimates the conditional mean log parameters. A prediction for the human species also includes residual heterogeneity, whereas uncertainty in the conditional mean does not. In either case, projection is strongest when the human predictor values lie within the range represented by the species data and when the modeled predictor-parameter relationship has a credible biological basis.

Joint Heterogeneity, Diagnostics, and Alternatives

Potency and slope should not automatically be analyzed as unrelated outcomes. With \(\psi_s=(\log EC_{50,s},\log h_s)^\mathsf{T}\), the off-diagonal element of \(\Omega\) captures whether residual deviations in the two parameters tend to occur together. Estimating that covariance can improve realistic joint human projections, but requires enough species and sufficiently informative curves. A correlation estimate from very few species is unstable, and a complex covariance structure can be poorly identified.

Useful checks begin with the inputs and the fitted parameter estimates: confirm common units and endpoint definitions, inspect whether each species' exposure range covers the informative part of its curve, and review the uncertainty and correlation in its fitted parameters. Then examine whether estimated species deviations and uncertainty are consistent with the assumed hierarchy, and whether the human covariate values are supported by the data range. Large uncertainty, highly correlated parameter estimates, or a prediction driven by extrapolation signal limited information. These checks identify reasons to be cautious; they do not establish that a model is biologically correct.

A fixed-effect analysis sets \(\tau^2=0\) and assumes that all species share the same true parameter after any specified covariate adjustment. It can be efficient if that assumption is credible, but understates uncertainty when meaningful heterogeneity remains. Naively pooling all individual observations into one curve is even more restrictive: it can confound species differences with exposure effects and give species with more observations disproportionate influence. A hierarchical model retains species-level structure, at the cost of estimating additional variance and covariance parameters.

The principal limitations are the number and choice of species, the quality and comparability of their data, and the biological validity of the predictors. A species-level association does not show that changing the predictor causes a parameter change. Moreover, transforming a mean log parameter gives a geometric-scale summary, not generally the arithmetic mean parameter; averaging nonlinear response curves also differs from evaluating a curve at average parameters. Human projections should therefore state which quantity is reported and which parameters were transported, estimated, or held fixed.